The Who's Who of AI

“Dropping a simple colored icon into a VLM's context can flip its coding strategy to 100% recursive—even when 71.5% of its thinking traces claim it's ignoring the image 👀 Wild study from Google DeepMind & Stanford on implicit visual steering: https://arxiv.o…”

“Research indicates vision-language models convey concepts with visuals but misalign with human interpretations. Simple images like hearts can greatly influence model behavior on unrelated tasks, revealing challenges in understanding AI's visual communicatio…”

“Access the updated Gemma 4 12B file: https://t.co/N8QOgLNc5C Quickly try Gemma using LiteRT-LM CLI: https://t.co/E4e6Quyu9D Learn more about building with LiteRT-LM: https://t.co/dmBCWFdbEB”

“🤖💎 Reachy on Wheels runs a 100% local, battery-powered robotics loop on LiteRT: Gemma 4 E4B handles voice chat + start/stop tool calls, while EmbeddingGemma 2 steers on an Arduino Ventuno Q from live camera input with zero cloud calls! https://developers.go…”

“Gemma 4 is running in Low Earth Orbit! 💎🛰️ Project Suncatcher ran the first ML inference on data-center TPUs in space: 4 TPUs in a ~1 kW solar envelope running Gemma 4 onboard and beaming a haiku to Earth: https://blog.google/innovation-and-ai/technology/re…”

“We launched a new home for Gemma developer docs at https://gemma.ai 📚💎 Includes a global model selector that adapts code snippets across sizes, one-click Markdown export on every page for local coding agents, and live on-device WebGPU demos in the browser.”

“🔍💎 Really neat agentic vision recipe from Meta's Muse Glimmer writeup! Instead of 1-shot GUI grounding on ScreenSpot-Pro, their 10-turn loop crops, zooms & overlays a grid. Running Gemma 4 31B through it hits 75.9% out of the box! https://research.meta.ai/s…”

“Launched Oct 1 with Planet Labs, the satellite took a prompt over Svalbard, ran Gemma 4 in orbit, and downlinked the response ahead of radiation and 2027 optical-cluster tests. System design: https://research.google/blog/exploring-a-space-based-scalable-ai-…”

“🔍🛠️ QUAN-Transcribe fine-tunes the top 2 USM conformer layers of EmbeddingGemma 2 with CTC in a single forward pass (~3,700x faster than real-time on H100), supporting batch and 0.48s streaming. Model weights: https://huggingface.co/qnguyen3/QUAN-Transcribe…”

“🎙️💎 Wild what you can do with multimodal embeddings: with 3h of fine-tuning on 3,000h of Vietnamese audio, Quan Nguyen (qnguyen3) used EmbeddingGemma 2 to top 8 Vietnamese ASR benchmarks with zero autoregressive decoder! https://github.com/qnguyen3/QUAN-Tra…”

“Ok - the fruits of my latest vibe AI engineering are ready: An all in-browser vector index (<4MB) + custom 4-layer EG2-compatible query embedder (~25MB) + custom WGSL runtime (KBs) of 16,878 Bluesky images, weighing in at a combined < 30MB INSTANT search as…”

“🔍🛠️ Pairing a compact E2B or E4B watchdog locally alongside your main coding or workflow agent is a clean pattern to validate tool arguments before execution. Weights on Hugging Face: https://huggingface.co/aisingapore/Gemma-SEA-LION-Guard-v2-31B”

“Neat browser-native build by Jakub Chmura at Software Mansion! 🖼️🔍⚡ They ported EmbeddingGemma 2 to ruNNtime (TypeScript WebGPU engine on TypeGPU) to search a local photo gallery in natural language 100% on the browser GPU. Repo: https://github.com/software…”

“👀💎 A Google engineer paired with Gemini 4 Argon in an iterative profiling and kernel-optimization loop to pinpoint the exact vLLM scheduling bottlenecks and land all 8 upstream PRs in days. Full PR list on GitHub: https://goo.gle/4zNtANl”

“🛠️💎 Bay Area builders: join our 100% on-device Gemma Hardware Hackathon at AGI House this Sat, Oct 24 with NVIDIA! 🤖 Hands-on with 15 Jetson Thor & 15 Orin Nano Super dev kits to build local edge AI & robotics with zero cloud: https://www.agihouse.org/event…”